2 # CUDA_SELECT_NVCC_ARCH_FLAGS(out_variable [target_CUDA_architectures])
3 # -- Selects GPU arch flags for nvcc based on target_CUDA_architectures
4 # target_CUDA_architectures : Auto | Common | All | LIST(ARCH_AND_PTX ...)
5 # - "Auto" detects local machine GPU compute arch at runtime.
6 # - "Common" and "All" cover common and entire subsets of architectures
7 # ARCH_AND_PTX : NAME | NUM.NUM | NUM.NUM(NUM.NUM) | NUM.NUM+PTX
8 # NAME: Fermi Kepler Maxwell Kepler+Tegra Kepler+Tesla Maxwell+Tegra Pascal Volta Turing Ampere
9 # NUM: Any number. Only those pairs are currently accepted by NVCC though:
10 # 2.0 2.1 3.0 3.2 3.5 3.7 5.0 5.2 5.3 6.0 6.2 7.0 7.2 7.5 8.0 8.6
11 # Returns LIST of flags to be added to CUDA_NVCC_FLAGS in ${out_variable}
12 # Additionally, sets ${out_variable}_readable to the resulting numeric list
14 # CUDA_SELECT_NVCC_ARCH_FLAGS(ARCH_FLAGS 3.0 3.5+PTX 5.2(5.0) Maxwell)
15 # LIST(APPEND CUDA_NVCC_FLAGS ${ARCH_FLAGS})
17 # More info on CUDA architectures: https://en.wikipedia.org/wiki/CUDA
20 if(CMAKE_CUDA_COMPILER_LOADED) # CUDA as a language
21 if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA"
22 AND CMAKE_CUDA_COMPILER_VERSION MATCHES "^([0-9]+\\.[0-9]+)")
23 set(CUDA_VERSION "${CMAKE_MATCH_1}")
27 # See: https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#gpu-feature-list
28 # Additions, deprecations, and removals can be found in the release notes:
29 # https://developer.nvidia.com/cuda-toolkit-archive
31 # The initial status here is for CUDA 7.0
32 set(CUDA_KNOWN_GPU_ARCHITECTURES "Fermi" "Kepler" "Maxwell" "Kepler+Tegra" "Kepler+Tesla" "Maxwell+Tegra")
33 set(CUDA_COMMON_GPU_ARCHITECTURES "2.0" "2.1" "3.0" "3.5" "5.0" "5.3")
34 set(CUDA_LIMIT_GPU_ARCHITECTURE "6.0")
35 set(CUDA_ALL_GPU_ARCHITECTURES "2.0" "2.1" "3.0" "3.2" "3.5" "3.7" "5.0" "5.2" "5.3")
36 set(_CUDA_MAX_COMMON_ARCHITECTURE "5.2+PTX")
39 if(CUDA_VERSION VERSION_GREATER_EQUAL "8.0")
40 list(APPEND CUDA_KNOWN_GPU_ARCHITECTURES "Pascal")
41 list(APPEND CUDA_COMMON_GPU_ARCHITECTURES "6.0" "6.1")
42 list(APPEND CUDA_ALL_GPU_ARCHITECTURES "6.0" "6.1" "6.2")
44 set(_CUDA_MAX_COMMON_ARCHITECTURE "6.2+PTX")
45 set(CUDA_LIMIT_GPU_ARCHITECTURE "7.0")
47 list(REMOVE_ITEM CUDA_COMMON_GPU_ARCHITECTURES "2.0" "2.1")
50 if(CUDA_VERSION VERSION_GREATER_EQUAL "9.0")
51 list(APPEND CUDA_KNOWN_GPU_ARCHITECTURES "Volta")
52 list(APPEND CUDA_COMMON_GPU_ARCHITECTURES "7.0")
53 list(APPEND CUDA_ALL_GPU_ARCHITECTURES "7.0" "7.2")
55 set(_CUDA_MAX_COMMON_ARCHITECTURE "7.2+PTX")
56 set(CUDA_LIMIT_GPU_ARCHITECTURE "8.0")
58 list(REMOVE_ITEM CUDA_KNOWN_GPU_ARCHITECTURES "Fermi")
59 list(REMOVE_ITEM CUDA_ALL_GPU_ARCHITECTURES "2.0" "2.1")
62 if(CUDA_VERSION VERSION_GREATER_EQUAL "10.0")
63 list(APPEND CUDA_KNOWN_GPU_ARCHITECTURES "Turing")
64 list(APPEND CUDA_COMMON_GPU_ARCHITECTURES "7.5")
65 list(APPEND CUDA_ALL_GPU_ARCHITECTURES "7.5")
67 set(_CUDA_MAX_COMMON_ARCHITECTURE "7.5+PTX")
68 set(CUDA_LIMIT_GPU_ARCHITECTURE "8.0")
70 list(REMOVE_ITEM CUDA_COMMON_GPU_ARCHITECTURES "3.0")
73 # https://docs.nvidia.com/cuda/archive/11.0/cuda-toolkit-release-notes/index.html#cuda-general-new-features
74 # https://docs.nvidia.com/cuda/archive/11.0/cuda-toolkit-release-notes/index.html#deprecated-features
75 if(CUDA_VERSION VERSION_GREATER_EQUAL "11.0")
76 list(APPEND CUDA_KNOWN_GPU_ARCHITECTURES "Ampere")
77 list(APPEND CUDA_COMMON_GPU_ARCHITECTURES "8.0")
78 list(APPEND CUDA_ALL_GPU_ARCHITECTURES "8.0")
80 set(_CUDA_MAX_COMMON_ARCHITECTURE "8.0+PTX")
81 set(CUDA_LIMIT_GPU_ARCHITECTURE "8.6")
83 list(REMOVE_ITEM CUDA_COMMON_GPU_ARCHITECTURES "3.5" "5.0")
84 list(REMOVE_ITEM CUDA_ALL_GPU_ARCHITECTURES "3.0" "3.2")
87 if(CUDA_VERSION VERSION_GREATER_EQUAL "11.1")
88 list(APPEND CUDA_COMMON_GPU_ARCHITECTURES "8.6")
89 list(APPEND CUDA_ALL_GPU_ARCHITECTURES "8.6")
91 set(_CUDA_MAX_COMMON_ARCHITECTURE "8.6+PTX")
92 set(CUDA_LIMIT_GPU_ARCHITECTURE "9.0")
95 list(APPEND CUDA_COMMON_GPU_ARCHITECTURES "${_CUDA_MAX_COMMON_ARCHITECTURE}")
97 # Check with: cmake -DCUDA_VERSION=7.0 -P select_compute_arch.cmake
98 if(DEFINED CMAKE_SCRIPT_MODE_FILE)
99 include(CMakePrintHelpers)
100 cmake_print_variables(CUDA_KNOWN_GPU_ARCHITECTURES)
101 cmake_print_variables(CUDA_COMMON_GPU_ARCHITECTURES)
102 cmake_print_variables(CUDA_LIMIT_GPU_ARCHITECTURE)
103 cmake_print_variables(CUDA_ALL_GPU_ARCHITECTURES)
107 ################################################################################################
108 # A function for automatic detection of GPUs installed (if autodetection is enabled)
110 # CUDA_DETECT_INSTALLED_GPUS(OUT_VARIABLE)
112 function(CUDA_DETECT_INSTALLED_GPUS OUT_VARIABLE)
113 if(NOT CUDA_GPU_DETECT_OUTPUT)
114 if(CMAKE_CUDA_COMPILER_LOADED) # CUDA as a language
115 set(file "${PROJECT_BINARY_DIR}/detect_cuda_compute_capabilities.cu")
117 set(file "${PROJECT_BINARY_DIR}/detect_cuda_compute_capabilities.cpp")
120 file(WRITE ${file} ""
121 "#include <cuda_runtime.h>\n"
122 "#include <cstdio>\n"
126 " if (cudaSuccess != cudaGetDeviceCount(&count)) return -1;\n"
127 " if (count == 0) return -1;\n"
128 " for (int device = 0; device < count; ++device)\n"
130 " cudaDeviceProp prop;\n"
131 " if (cudaSuccess == cudaGetDeviceProperties(&prop, device))\n"
132 " std::printf(\"%d.%d \", prop.major, prop.minor);\n"
137 if(CMAKE_CUDA_COMPILER_LOADED) # CUDA as a language
138 try_run(run_result compile_result SOURCES ${file}
139 RUN_OUTPUT_VARIABLE compute_capabilities)
141 try_run(run_result compile_result SOURCES ${file}
142 CMAKE_FLAGS "-DINCLUDE_DIRECTORIES=${CUDA_INCLUDE_DIRS}"
143 LINK_LIBRARIES ${CUDA_LIBRARIES}
144 RUN_OUTPUT_VARIABLE compute_capabilities)
147 # Filter unrelated content out of the output.
148 string(REGEX MATCHALL "[0-9]+\\.[0-9]+" compute_capabilities "${compute_capabilities}")
150 if(run_result EQUAL 0)
151 string(REPLACE "2.1" "2.1(2.0)" compute_capabilities "${compute_capabilities}")
152 set(CUDA_GPU_DETECT_OUTPUT ${compute_capabilities}
153 CACHE INTERNAL "Returned GPU architectures from detect_gpus tool" FORCE)
157 if(NOT CUDA_GPU_DETECT_OUTPUT)
158 message(STATUS "Automatic GPU detection failed. Building for common architectures.")
159 set(${OUT_VARIABLE} ${CUDA_COMMON_GPU_ARCHITECTURES} PARENT_SCOPE)
161 # Filter based on CUDA version supported archs
162 set(CUDA_GPU_DETECT_OUTPUT_FILTERED "")
163 separate_arguments(CUDA_GPU_DETECT_OUTPUT)
164 foreach(ITEM IN ITEMS ${CUDA_GPU_DETECT_OUTPUT})
165 if(CUDA_LIMIT_GPU_ARCHITECTURE AND ITEM VERSION_GREATER_EQUAL CUDA_LIMIT_GPU_ARCHITECTURE)
166 list(GET CUDA_COMMON_GPU_ARCHITECTURES -1 NEWITEM)
167 string(APPEND CUDA_GPU_DETECT_OUTPUT_FILTERED " ${NEWITEM}")
169 string(APPEND CUDA_GPU_DETECT_OUTPUT_FILTERED " ${ITEM}")
173 set(${OUT_VARIABLE} ${CUDA_GPU_DETECT_OUTPUT_FILTERED} PARENT_SCOPE)
178 ################################################################################################
179 # Function for selecting GPU arch flags for nvcc based on CUDA architectures from parameter list
181 # SELECT_NVCC_ARCH_FLAGS(out_variable [list of CUDA compute archs])
182 function(CUDA_SELECT_NVCC_ARCH_FLAGS out_variable)
183 set(CUDA_ARCH_LIST "${ARGN}")
185 if("X${CUDA_ARCH_LIST}" STREQUAL "X" )
186 set(CUDA_ARCH_LIST "Auto")
192 if("${CUDA_ARCH_LIST}" STREQUAL "All")
193 set(CUDA_ARCH_LIST ${CUDA_KNOWN_GPU_ARCHITECTURES})
194 elseif("${CUDA_ARCH_LIST}" STREQUAL "Common")
195 set(CUDA_ARCH_LIST ${CUDA_COMMON_GPU_ARCHITECTURES})
196 elseif("${CUDA_ARCH_LIST}" STREQUAL "Auto")
197 CUDA_DETECT_INSTALLED_GPUS(CUDA_ARCH_LIST)
198 message(STATUS "Autodetected CUDA architecture(s): ${CUDA_ARCH_LIST}")
201 # Now process the list and look for names
202 string(REGEX REPLACE "[ \t]+" ";" CUDA_ARCH_LIST "${CUDA_ARCH_LIST}")
203 list(REMOVE_DUPLICATES CUDA_ARCH_LIST)
204 foreach(arch_name ${CUDA_ARCH_LIST})
208 # Check to see if we are compiling PTX
209 if(arch_name MATCHES "(.*)\\+PTX$")
211 set(arch_name ${CMAKE_MATCH_1})
213 if(arch_name MATCHES "^([0-9]\\.[0-9](\\([0-9]\\.[0-9]\\))?)$")
214 set(arch_bin ${CMAKE_MATCH_1})
215 set(arch_ptx ${arch_bin})
217 # Look for it in our list of known architectures
218 if(${arch_name} STREQUAL "Fermi")
219 set(arch_bin 2.0 "2.1(2.0)")
220 elseif(${arch_name} STREQUAL "Kepler+Tegra")
222 elseif(${arch_name} STREQUAL "Kepler+Tesla")
224 elseif(${arch_name} STREQUAL "Kepler")
225 set(arch_bin 3.0 3.5)
227 elseif(${arch_name} STREQUAL "Maxwell+Tegra")
229 elseif(${arch_name} STREQUAL "Maxwell")
230 set(arch_bin 5.0 5.2)
232 elseif(${arch_name} STREQUAL "Pascal")
233 set(arch_bin 6.0 6.1)
235 elseif(${arch_name} STREQUAL "Volta")
236 set(arch_bin 7.0 7.0)
238 elseif(${arch_name} STREQUAL "Turing")
241 elseif(${arch_name} STREQUAL "Ampere")
245 message(SEND_ERROR "Unknown CUDA Architecture Name ${arch_name} in CUDA_SELECT_NVCC_ARCH_FLAGS")
249 message(SEND_ERROR "arch_bin wasn't set for some reason")
251 list(APPEND cuda_arch_bin ${arch_bin})
254 set(arch_ptx ${arch_bin})
256 list(APPEND cuda_arch_ptx ${arch_ptx})
260 # remove dots and convert to lists
261 string(REGEX REPLACE "\\." "" cuda_arch_bin "${cuda_arch_bin}")
262 string(REGEX REPLACE "\\." "" cuda_arch_ptx "${cuda_arch_ptx}")
263 string(REGEX MATCHALL "[0-9()]+" cuda_arch_bin "${cuda_arch_bin}")
264 string(REGEX MATCHALL "[0-9]+" cuda_arch_ptx "${cuda_arch_ptx}")
267 list(REMOVE_DUPLICATES cuda_arch_bin)
270 list(REMOVE_DUPLICATES cuda_arch_ptx)
274 set(nvcc_archs_readable "")
276 # Tell NVCC to add binaries for the specified GPUs
277 foreach(arch ${cuda_arch_bin})
278 if(arch MATCHES "([0-9]+)\\(([0-9]+)\\)")
279 # User explicitly specified ARCH for the concrete CODE
280 list(APPEND nvcc_flags -gencode arch=compute_${CMAKE_MATCH_2},code=sm_${CMAKE_MATCH_1})
281 list(APPEND nvcc_archs_readable sm_${CMAKE_MATCH_1})
283 # User didn't explicitly specify ARCH for the concrete CODE, we assume ARCH=CODE
284 list(APPEND nvcc_flags -gencode arch=compute_${arch},code=sm_${arch})
285 list(APPEND nvcc_archs_readable sm_${arch})
289 # Tell NVCC to add PTX intermediate code for the specified architectures
290 foreach(arch ${cuda_arch_ptx})
291 list(APPEND nvcc_flags -gencode arch=compute_${arch},code=compute_${arch})
292 list(APPEND nvcc_archs_readable compute_${arch})
295 string(REPLACE ";" " " nvcc_archs_readable "${nvcc_archs_readable}")
296 set(${out_variable} ${nvcc_flags} PARENT_SCOPE)
297 set(${out_variable}_readable ${nvcc_archs_readable} PARENT_SCOPE)